Abstract

Predicting drug solubility remains a major challenge in the pharmaceutical industry given its central role in formulation development and bioavailability. In this context, the Conductor-like Screening Model for Real Solvents (COSMO-RS) is widely employed to estimate solvation and solubility properties, yet the accuracy of these predictions can strongly depend on the molecular conformations provided as the input. Although this issue has been addressed in previous studies for specific molecules, no systematic evaluation of the influence of conformer-generation methods on the quality of COSMO-RS solubility predictions has been reported using large and chemically diverse data sets. Here, we assess four representative conformer-generation workflows, ranging from simple RDKit-based approaches to more elaborate semiempirical and density functional theory (DFT)-based pipelines, using a curated data set comprising 926 experimental solubility values and 405 free energies of solvation. Despite substantial methodological differences and computational cost, all workflows yield comparable accuracy across the full data set, with mean absolute errors close to one logarithmic unit, and all tend to underperform in the low-solubility regime, which remains a persistent challenge in computational chemistry. We further examined the use of conformer ensembles for representative molecules and compared their performance with solvent-aware single-conformer pipelines. For the cases analyzed, ensemble averaging provides moderate improvements but does not systematically outperform solvent-aware lowest-energy conformer selection. Notably, incorporating solvent information into the lowest-energy conformer identification stage can improve the accuracy for molecules capable of establishing intramolecular interactions, such as hydrogen bonds. Overall, our results provide practical guidance for selecting conformer-generation strategies in high-throughput COSMO-RS applications and highlight solvent-aware workflows as a promising direction to improve the predictive robustness without compromising computational efficiency.

Read the full article at Pubs.ACS.com